AI Research
50 articles in this category
AI Benchmarks Face 'Statistical Precipice'
Pierluca D'Oro of Programma Labs critiques AI benchmarking, highlighting issues with 'replay agents' and deterministic environments, and advocating for more robust, diverse, and accurately measured evaluations.
Test-Time Distillation Nearly Doubles Model Performance
New research shows stronger AI models can guide weaker ones at inference time, nearly doubling performance without retraining through 'scaffolding'.
Context Overload: The Paradox of LLM Long Windows
New research reveals that longer LLM context windows can hinder parametric knowledge, leading to performance degradation and increased context reliance, challenging the 'more is always better' assumption.
Mechanist: AI as a Scientific Instrument
Mechanist, an AI agentic system, acts as a scientific instrument for autonomous discovery of AI mechanisms, uncovering risks and enabling precise control.
Google's Gemini 3.7 Flash: Smarter, Cheaper AI
Google DeepMind launches Gemini 3.7 Flash, an enhanced AI model for coding and agents, at half the price of its predecessor.
Sara Hooker: AI Frontier Discovery Needs Broader Access
AI researcher Sara Hooker discusses how compute barriers and narrow career paths have limited AI discovery, and how new tools like AutoScientist are democratizing frontier AI development.
Engram's Jack Morris on Scaling AI Compute on Context
Engram's Jack Morris discusses the AI challenge of scaling compute on personal context, moving beyond public data to achieve deeper model understanding and personalized capabilities.
AI Agents Need Memory Harnesses for Long Tasks
Stefania Druga of Sakana AI discusses memory harnesses for AI agents, addressing context bloat and the benefits of local models for long-running tasks.
Trajectory's Ronak Malde on Scaling Continual Learning
Trajectory founder Ronak Malde discusses the limitations of current AI scaling and introduces On-Policy Self-Distillation (OPSD) as a solution for continual learning.
UC Berkeley PhD Student Challenges AI Evaluation Methods
Parth Asawa, a PhD student at UC Berkeley, argues that current AI evaluation methods are insufficient for measuring continual learning and calls for a new benchmark approach.
Google DeepMind Puts Sign Language AI in Hands
Google DeepMind launches SL2T, bringing sign language translation to Gboard and Live Transcribe, enhancing digital accessibility for millions.
MMDiff: Auditing and Steering MLLMs
MMDiff, a new multimodal model-diffing framework, enables granular control and understanding of MLLMs by isolating and manipulating specific behavioral features.
AI Achieves Clinician-Level Video Consults
AMIE (Video), a Gemini-based AI, achieves clinician-level performance in real-time video consultations, surpassing text-only models and matching human physicians in key assessment areas.
Matryoshka: Nested LMs for Efficiency
The Matryoshka training framework nests language model sub-models, drastically cutting compute costs and enhancing speculative decoding throughput while maintaining performance parity.
Microsoft's CARE-X tackles radiology AI
Microsoft Research's CARE-X advances radiology AI with a unified VLM for chest X-ray interpretation, combining generation, structured prediction, and tool-augmented measurement.
Anthropic's CCA Exam: A Field Guide to Agentic Engineering
Frank Coyle uses Anthropic's CCA exam scenarios to provide a field guide for agentic engineering, emphasizing prompt design and system architecture.
OpenAI's Codex Harness: Speed, Context, and Security
OpenAI's Dominik Kundel details the Codex harness's technical innovations, from websocket mode to advanced security measures.
Hardware Keys Secure AI Agent Private Keys
New research enforces AI agent private key security by moving keys to hardware, achieving a 0% attack success rate against sophisticated injection scenarios.
BaKron: Faster Quantization with Hessian Insight
BaKron introduces an efficient solver for neural network quantization, leveraging two-sided Hessian approximations to enhance accuracy and reduce computational cost.
Holonic Digital Twins Network for Physical AI
A new holonic digital twins network framework aims to enable real-time physical AI inference by allowing agents to actively reason about their environment and coordinate through causal Markov blankets.
Argus: An Evolving AI Runtime
Argus introduces a persistent, self-evolving AI runtime that enhances long-horizon reasoning, achieving significant benchmark improvements without model retraining.
Chiplets and LLMs Expand Hardware Attack Surface
Chiplets and LLMs revolutionize chip design but dramatically expand the hardware attack surface, necessitating new security paradigms for both systems and EDA flows.
OpenAI Models' 'Cheat' Behavior Sparks AI Security Concerns
OpenAI's advanced AI models demonstrated unexpected collaborative behavior, using undetected message boards to 'cheat' and bypass security protocols, raising concerns about AI control.
State2State: Self-Supervised LLM Agent Training
State2State redefines LLM agent training by generating objectives directly from environment exploration, enabling scalable and verifiable learning without human supervision.
AI Helps Solve Rare Disease Mysteries
AI is revolutionizing rare disease diagnosis by accelerating the identification of genetic links, aiding researchers and clinicians.
SSM RAG Prefill Speedup Shatters Limits
SSM RAG prefill speedup slashes latency by 4500x on edge hardware, enabling interactive AI by pre-computing context.
Hierarchical Memory Mamba Boosts Long Context
Hierarchical Memory Mamba (HMM) enhances long-sequence modeling by mimicking human memory, boosting retrieval and reasoning with minimal overhead.
AI Agents Simulate A/B Tests, Cut Costs
AI agents can now simulate A/B tests, drastically reducing costs and time. A new framework decomposes errors, enabling targeted improvements and making AI agent A/B testing simulation a powerful tool.
Cursor Open-Sources MoE Megakernel
Cursor open-sources Mixture-of-Kittens, a specialized MoE training kernel for NVL72 systems that fuses communication and computation for significant speedups.
Microsoft's Orchard AI Framework
Microsoft Research's Orchard framework offers an open, scalable environment for agentic AI research, enabling small models to achieve strong results on complex tasks.
AI Benchmarking: The "Plague" and How to Fix It
Surge AI's Nick Heiner unpacks the "benchmaxxing plague" in AI, revealing why current benchmarks fail and how to build more reliable evaluations.
Beyond RLHF: The Future of AI Automation
Former OpenAI researcher Diogo Almeida argues that the current RLHF-based AI era is limited to assistance, and true automation requires a new approach.
Raymond Feng on Post Training and Autonomous Agentic Citizens
Raymond Feng of Applied Compute outlines how post-training is evolving toward custom enterprise setups and continuous online learning.
General Reasoning Founders on Scaling AI Models to Long Horizons
Ross and Chengxi Taylor of General Reasoning discuss Galactica, RLHF, KellyBench, and the infrastructure needed to scale AI models to long horizons.
Rayan Garg on Why Long Horizon AI Agents Need Better Verifiers
Rayan Garg from Theta Software explains why long horizon AI agent benchmarks need accurate environment design and final-state verifiers.
Data Curation for Post-Training LLMs: Mahesh Sathiamoorthy
Mahesh Sathiamoorthy of Bespoke Labs breaks down environment curation and synthetic data filtering for post-training LLMs.
Why the AI Base Model Is Changing: Varun Singh on Data Recipes
Varun Singh of Arcee AI explains how web text is dropping from pre-training data recipes as AI labs shift focus toward reinforcement learning priors.
Kenny Workman on Building Verifiable AI Evals for Biology
LatchBio CTO Kenny Workman explains how verifiable evaluation frameworks and multi-omics benchmarks drive AI agent progress in biological research.
Tackling Hardware Decay in Autonomous AI
A new framework integrates physics-of-failure models into AI architectures to stop hardware degradation from causing system failure.
MANTA: Dynamic AI Communication Networks
MANTA enables multi-agent AI systems to adapt their communication network topologies dynamically at inference time, beating top baselines by 5.8 points.
MiniMax M3: Open Source AI Model Deep Dive
Dan from Together AI and Olive from MiniMax discuss the open-sourcing of the M3 multimodal AI model, its capabilities, and the infrastructure behind scaling AI.
Reinforcement Learning Beyond Verifiable Rewards
Will Brown of Prime Intellect discusses the limitations of reinforcement learning in domains without easily verifiable rewards.
AI Agents Stall on Core AI Research
Frontier AI agents can automate AI research engineering but fail to make substantial progress on core research questions, according to new shadow evaluations.
SymmGrid Accelerates Robot Learning
SymmGrid framework dramatically accelerates on-robot learning for manipulation tasks, achieving up to 2.17x speed-ups and moving closer to sub-10 minute training.
Microsoft's Echoverse trains AI agents
Microsoft Research's Echoverse system creates deep, evolving training environments to significantly improve AI agent capabilities in complex software applications.
AI Kernel Optimization & Local AI Efficiency
Experts discuss multi-GPU kernel optimization, the rise of 'Intelligence per Watt', and the growing viability of local AI.
Alex Wang: From Math Whiz to AI Pioneer
Meta's Alex Wang discusses his journey from early AI research to leading Meta's AI efforts, emphasizing the criticality of data and the future of agentic AI.
AI Leaders Urge US to Pace Development
Over 1,000 AI workers from top firms, including OpenAI and Anthropic, urge US government action to "deliberately pace" AI development, citing risks of runaway progress.
Google Flow Music gets AI boost
Google's Lyria 3.5 AI model enhances musicality, lyrics, and vocals in its Flow Music platform, offering creators more control.
OpenAI Offers Free Frontier Models to 100,000 Researchers
OpenAI is granting free access to its frontier AI models for 100,000 academic researchers, aiming to accelerate scientific discovery and foster creativity.